自动化元数据提取以改进检索 + 合成¶

在本教程中，我们将向您展示如何执行自动化元数据提取以获取更好的检索结果。我们使用两种提取器：一种是 QuestionAnsweredExtractor，它从一段文本中生成问答对；另一种是 SummaryExtractor，它不仅提取当前文本的摘要，还提取相邻文本的摘要。

我们证明，这使得“块梦想”（chunk dreaming）成为可能——每个单独的块可以拥有更“整体的”细节，从而在给定检索结果的情况下提高答案质量。

我们的数据源取自 Eugene Yan 关于 LLM 模式的流行文章：https://eugeneyan.com/writing/llm-patterns/

设置¶

如果您正在 Colab 上打开此 Notebook，您可能需要安装 LlamaIndex 🦙。

输入 [ ]

%pip install llama-index-llms-openai
%pip install llama-index-readers-web

%pip install llama-index-llms-openai %pip install llama-index-readers-web

输入 [ ]

!pip install llama-index

!pip install llama-index

输入 [ ]

import nest_asyncio

nest_asyncio.apply()

import os
import openai

import nest_asyncio nest_asyncio.apply() import os import openai

输入 [ ]

# OPTIONAL: setup W&B callback handling for tracing
from llama_index.core import set_global_handler

set_global_handler("wandb", run_args={"project": "llamaindex"})

# 可选：设置 W&B 回调处理以进行追踪 from llama_index.core import set_global_handler set_global_handler("wandb", run_args={"project": "llamaindex"})

输入 [ ]

os.environ["OPENAI_API_KEY"] = "sk-..."
openai.api_key = os.environ["OPENAI_API_KEY"]

os.environ["OPENAI_API_KEY"] = "sk-..." openai.api_key = os.environ["OPENAI_API_KEY"]

定义元数据提取器¶

在此处，我们定义元数据提取器。我们定义两种变体：

• metadata_extractor_1 仅包含 QuestionsAnsweredExtractor
• metadata_extractor_2 包含 QuestionsAnsweredExtractor 和 SummaryExtractor

输入 [ ]

from llama_index.llms.openai import OpenAI
from llama_index.core.schema import MetadataMode

from llama_index.llms.openai import OpenAI from llama_index.core.schema import MetadataMode

输入 [ ]

llm = OpenAI(temperature=0.1, model="gpt-3.5-turbo", max_tokens=512)

llm = OpenAI(temperature=0.1, model="gpt-3.5-turbo", max_tokens=512)

我们还展示了如何实例化 SummaryExtractor 和 QuestionsAnsweredExtractor。

输入 [ ]

from llama_index.core.node_parser import TokenTextSplitter
from llama_index.core.extractors import (
    SummaryExtractor,
    QuestionsAnsweredExtractor,
)

node_parser = TokenTextSplitter(
    separator=" ", chunk_size=256, chunk_overlap=128
)


extractors_1 = [
    QuestionsAnsweredExtractor(
        questions=3, llm=llm, metadata_mode=MetadataMode.EMBED
    ),
]

extractors_2 = [
    SummaryExtractor(summaries=["prev", "self", "next"], llm=llm),
    QuestionsAnsweredExtractor(
        questions=3, llm=llm, metadata_mode=MetadataMode.EMBED
    ),
]

from llama_index.core.node_parser import TokenTextSplitter from llama_index.core.extractors import ( SummaryExtractor, QuestionsAnsweredExtractor, ) node_parser = TokenTextSplitter( separator=" ", chunk_size=256, chunk_overlap=128 ) extractors_1 = [ QuestionsAnsweredExtractor( questions=3, llm=llm, metadata_mode=MetadataMode.EMBED ), ] extractors_2 = [ SummaryExtractor(summaries=["prev", "self", "next"], llm=llm), QuestionsAnsweredExtractor( questions=3, llm=llm, metadata_mode=MetadataMode.EMBED ), ]

加载数据，运行提取器¶

我们使用 LlamaHub SimpleWebPageReader 加载 Eugene 的文章 (https://eugeneyan.com/writing/llm-patterns/)。

然后我们运行我们的提取器。

输入 [ ]

from llama_index.core import SimpleDirectoryReader

from llama_index.core import SimpleDirectoryReader

输入 [ ]

# load in blog

from llama_index.readers.web import SimpleWebPageReader

reader = SimpleWebPageReader(html_to_text=True)
docs = reader.load_data(urls=["https://eugeneyan.com/writing/llm-patterns/"])

# 加载博客 from llama_index.readers.web import SimpleWebPageReader reader = SimpleWebPageReader(html_to_text=True) docs = reader.load_data(urls=["https://eugeneyan.com/writing/llm-patterns/"])

输入 [ ]

print(docs[0].get_content())

print(docs[0].get_content())

输入 [ ]

orig_nodes = node_parser.get_nodes_from_documents(docs)

orig_nodes = node_parser.get_nodes_from_documents(docs)

输入 [ ]

# take just the first 8 nodes for testing
nodes = orig_nodes[20:28]

# 只取前 8 个节点用于测试 nodes = orig_nodes[20:28]

输入 [ ]

print(nodes[3].get_content(metadata_mode="all"))

print(nodes[3].get_content(metadata_mode="all"))

is to measure the distance that words would
have to move to convert one sequence to another.

However, there are several pitfalls to using these conventional benchmarks and
metrics.

First, there’s **poor correlation between these metrics and human judgments.**
BLEU, ROUGE, and others have had [negative correlation with how humans
evaluate fluency](https://arxiv.org/abs/2008.12009). They also showed moderate
to less correlation with human adequacy scores. In particular, BLEU and ROUGE
have [low correlation with tasks that require creativity and
diversity](https://arxiv.org/abs/2303.16634).

Second, these metrics often have **poor adaptability to a wider variety of
tasks**. Adopting a metric proposed for one task to another is not always
prudent. For example, exact match metrics such as BLEU and ROUGE are a poor
fit for tasks like abstractive summarization or dialogue. Since they’re based
on n-gram overlap between output and reference, they don’t make sense for a
dialogue task where a wide variety

运行元数据提取器¶

输入 [ ]

from llama_index.core.ingestion import IngestionPipeline

# process nodes with metadata extractors
pipeline = IngestionPipeline(transformations=[node_parser, *extractors_1])

nodes_1 = pipeline.run(nodes=nodes, in_place=False, show_progress=True)

from llama_index.core.ingestion import IngestionPipeline # 使用元数据提取器处理节点 pipeline = IngestionPipeline(transformations=[node_parser, *extractors_1]) nodes_1 = pipeline.run(nodes=nodes, in_place=False, show_progress=True)

Parsing documents into nodes:   0%|          | 0/8 [00:00<?, ?it/s]

Extracting questions:   0%|          | 0/8 [00:00<?, ?it/s]

输入 [ ]

print(nodes_1[3].get_content(metadata_mode="all"))

print(nodes_1[3].get_content(metadata_mode="all"))

[Excerpt from document]
questions_this_excerpt_can_answer: 1. What is the correlation between conventional metrics like BLEU and ROUGE and human judgments in evaluating fluency and adequacy in natural language processing tasks?
2. How do conventional metrics like BLEU and ROUGE perform in tasks that require creativity and diversity?
3. Why are exact match metrics like BLEU and ROUGE not suitable for tasks like abstractive summarization or dialogue in natural language processing?
Excerpt:
-----
is to measure the distance that words would
have to move to convert one sequence to another.

However, there are several pitfalls to using these conventional benchmarks and
metrics.

First, there’s **poor correlation between these metrics and human judgments.**
BLEU, ROUGE, and others have had [negative correlation with how humans
evaluate fluency](https://arxiv.org/abs/2008.12009). They also showed moderate
to less correlation with human adequacy scores. In particular, BLEU and ROUGE
have [low correlation with tasks that require creativity and
diversity](https://arxiv.org/abs/2303.16634).

Second, these metrics often have **poor adaptability to a wider variety of
tasks**. Adopting a metric proposed for one task to another is not always
prudent. For example, exact match metrics such as BLEU and ROUGE are a poor
fit for tasks like abstractive summarization or dialogue. Since they’re based
on n-gram overlap between output and reference, they don’t make sense for a
dialogue task where a wide variety
-----

输入 [ ]

# 2nd pass: run summaries, and then metadata extractor

# process nodes with metadata extractor
pipeline = IngestionPipeline(transformations=[node_parser, *extractors_2])

nodes_2 = pipeline.run(nodes=nodes, in_place=False, show_progress=True)

# 第二遍：运行摘要，然后运行元数据提取器 # 使用元数据提取器处理节点 pipeline = IngestionPipeline(transformations=[node_parser, *extractors_2]) nodes_2 = pipeline.run(nodes=nodes, in_place=False, show_progress=True)

Parsing documents into nodes:   0%|          | 0/8 [00:00<?, ?it/s]

Extracting summaries:   0%|          | 0/8 [00:00<?, ?it/s]

Extracting questions:   0%|          | 0/8 [00:00<?, ?it/s]

可视化一些样本数据¶

输入 [ ]

print(nodes_2[3].get_content(metadata_mode="all"))

print(nodes_2[3].get_content(metadata_mode="all"))

[Excerpt from document]
prev_section_summary: The section discusses the comparison between BERTScore and MoverScore, two metrics used to evaluate the quality of text generation models. MoverScore is described as a metric that measures the effort required to transform one text sequence into another by mapping semantically related words. The section also highlights the limitations of conventional benchmarks and metrics, such as poor correlation with human judgments and low correlation with tasks requiring creativity.
next_section_summary: The section discusses the limitations of current evaluation metrics in natural language processing tasks. It highlights three main issues: lack of creativity and diversity in metrics, poor adaptability to different tasks, and poor reproducibility. The section mentions specific metrics like BLEU and ROUGE, and also references studies that have reported high variance in metric scores.
section_summary: The section discusses the limitations of conventional benchmarks and metrics used to measure the distance between word sequences. It highlights two main issues: the poor correlation between these metrics and human judgments, and their limited adaptability to different tasks. The section mentions specific metrics like BLEU and ROUGE, which have been found to have low correlation with human evaluations of fluency, adequacy, creativity, and diversity. It also points out that metrics based on n-gram overlap, such as BLEU and ROUGE, are not suitable for tasks like abstractive summarization or dialogue.
questions_this_excerpt_can_answer: 1. What are the limitations of conventional benchmarks and metrics in measuring the distance between word sequences?
2. How do metrics like BLEU and ROUGE correlate with human judgments in terms of fluency, adequacy, creativity, and diversity?
3. Why are metrics based on n-gram overlap, such as BLEU and ROUGE, not suitable for tasks like abstractive summarization or dialogue?
Excerpt:
-----
is to measure the distance that words would
have to move to convert one sequence to another.

However, there are several pitfalls to using these conventional benchmarks and
metrics.

First, there’s **poor correlation between these metrics and human judgments.**
BLEU, ROUGE, and others have had [negative correlation with how humans
evaluate fluency](https://arxiv.org/abs/2008.12009). They also showed moderate
to less correlation with human adequacy scores. In particular, BLEU and ROUGE
have [low correlation with tasks that require creativity and
diversity](https://arxiv.org/abs/2303.16634).

Second, these metrics often have **poor adaptability to a wider variety of
tasks**. Adopting a metric proposed for one task to another is not always
prudent. For example, exact match metrics such as BLEU and ROUGE are a poor
fit for tasks like abstractive summarization or dialogue. Since they’re based
on n-gram overlap between output and reference, they don’t make sense for a
dialogue task where a wide variety
-----

输入 [ ]

print(nodes_2[1].get_content(metadata_mode="all"))

print(nodes_2[1].get_content(metadata_mode="all"))

[Excerpt from document]
prev_section_summary: The section discusses the F_{BERT} formula used in BERTScore and highlights the advantages of BERTScore over simpler metrics like BLEU and ROUGE. It also introduces MoverScore, another metric that uses contextualized embeddings but allows for many-to-one matching. The key topics are BERTScore, MoverScore, and the differences between them.
next_section_summary: The section discusses the comparison between BERTScore and MoverScore, two metrics used to evaluate the quality of text generation models. MoverScore is described as a metric that measures the effort required to transform one text sequence into another by mapping semantically related words. The section also highlights the limitations of conventional benchmarks and metrics, such as poor correlation with human judgments and low correlation with tasks requiring creativity.
section_summary: The key topics of this section are BERTScore and MoverScore, which are methods used to compute the similarity between generated output and reference in tasks like image captioning and machine translation. BERTScore uses one-to-one matching of tokens, while MoverScore allows for many-to-one matching. MoverScore solves an optimization problem to measure the distance that words would have to move to convert one sequence to another.
questions_this_excerpt_can_answer: 1. What is the main difference between BERTScore and MoverScore?
2. How does MoverScore allow for many-to-one matching of tokens?
3. What problem does MoverScore solve to measure the distance between two sequences?
Excerpt:
-----
to have better correlation for tasks
such as image captioning and machine translation.

**[MoverScore](https://arxiv.org/abs/1909.02622)** also uses contextualized
embeddings to compute the distance between tokens in the generated output and
reference. But unlike BERTScore, which is based on one-to-one matching (or
“hard alignment”) of tokens, MoverScore allows for many-to-one matching (or
“soft alignment”).

![BERTScore \(left\) vs. MoverScore \(right\)](/assets/mover-score.jpg)

BERTScore (left) vs. MoverScore (right;
[source](https://arxiv.org/abs/1909.02622))

MoverScore enables the mapping of semantically related words in one sequence
to their counterparts in another sequence. It does this by solving a
constrained optimization problem that finds the minimum effort to transform
one text into another. The idea is to measure the distance that words would
have to move to convert one sequence to another.

However, there
-----

设置 RAG 查询引擎，比较结果！¶

我们在三种节点变体之上设置了 3 个索引/查询引擎。

输入 [ ]

from llama_index.core import VectorStoreIndex
from llama_index.core.response.notebook_utils import (
    display_source_node,
    display_response,
)

from llama_index.core import VectorStoreIndex from llama_index.core.response.notebook_utils import ( display_source_node, display_response, )

输入 [ ]

# try out different query engines

# index0 = VectorStoreIndex(orig_nodes)
# index1 = VectorStoreIndex(nodes_1 + orig_nodes[8:])
# index2 = VectorStoreIndex(nodes_2 + orig_nodes[8:])

index0 = VectorStoreIndex(orig_nodes)
index1 = VectorStoreIndex(orig_nodes[:20] + nodes_1 + orig_nodes[28:])
index2 = VectorStoreIndex(orig_nodes[:20] + nodes_2 + orig_nodes[28:])

# 尝试不同的查询引擎 # index0 = VectorStoreIndex(orig_nodes) # index1 = VectorStoreIndex(nodes_1 + orig_nodes[8:]) # index2 = VectorStoreIndex(nodes_2 + orig_nodes[8:]) index0 = VectorStoreIndex(orig_nodes) index1 = VectorStoreIndex(orig_nodes[:20] + nodes_1 + orig_nodes[28:]) index2 = VectorStoreIndex(orig_nodes[:20] + nodes_2 + orig_nodes[28:])

输入 [ ]

query_engine0 = index0.as_query_engine(similarity_top_k=1)
query_engine1 = index1.as_query_engine(similarity_top_k=1)
query_engine2 = index2.as_query_engine(similarity_top_k=1)

query_engine0 = index0.as_query_engine(similarity_top_k=1) query_engine1 = index1.as_query_engine(similarity_top_k=1) query_engine2 = index2.as_query_engine(similarity_top_k=1)

尝试一些问题¶

在这个问题中，我们看到朴素的响应 response0 只提到了 BLEU 和 ROUGE，并且缺乏关于其他指标的上下文。

另一方面，response2 在其上下文中包含了所有指标。

输入 [ ]

# query_str = "In the original RAG paper, can you describe the two main approaches for generation and compare them?"
query_str = (
    "Can you describe metrics for evaluating text generation quality, compare"
    " them, and tell me about their downsides"
)

response0 = query_engine0.query(query_str)
response1 = query_engine1.query(query_str)
response2 = query_engine2.query(query_str)

# query_str = “在最初的 RAG 论文中，你能描述两种主要的生成方法并进行比较吗？” query_str = ( “你能描述评估文本生成质量的指标，比较” “它们，并告诉我它们的缺点” )

输入 [ ]

display_response(
    response0, source_length=1000, show_source=True, show_source_metadata=True
)

display_response( response0, source_length=1000, show_source=True, show_source_metadata=True )

输入 [ ]

print(response0.source_nodes[0].node.get_content())

print(response0.source_nodes[0].node.get_content())

require creativity and
diversity](https://arxiv.org/abs/2303.16634).

Second, these metrics often have **poor adaptability to a wider variety of
tasks**. Adopting a metric proposed for one task to another is not always
prudent. For example, exact match metrics such as BLEU and ROUGE are a poor
fit for tasks like abstractive summarization or dialogue. Since they’re based
on n-gram overlap between output and reference, they don’t make sense for a
dialogue task where a wide variety of responses are possible. An output can
have zero n-gram overlap with the reference but yet be a good response.

Third, these metrics have **poor reproducibility**. Even for the same metric,
[high variance is reported across different
studies](https://arxiv.org/abs/2008.12009), possibly due to variations in
human judgment collection or metric parameter settings. Another study of
[ROUGE scores](https://aclanthology.org/2023.acl-long.107/) across 2,000
studies found that scores were hard

输入 [ ]

display_response(
    response1, source_length=1000, show_source=True, show_source_metadata=True
)

display_response( response1, source_length=1000, show_source=True, show_source_metadata=True )

输入 [ ]

display_response(
    response2, source_length=1000, show_source=True, show_source_metadata=True
)

display_response( response2, source_length=1000, show_source=True, show_source_metadata=True )

在下一个问题中，我们询问 BERTScore/MoverScore。

响应是相似的。但 response2 提供了比 response0 略多的细节，因为它在元数据中包含更多关于 MoverScore 的信息。

输入 [ ]

# query_str = "What are some reproducibility issues with the ROUGE metric? Give some details related to benchmarks and also describe other ROUGE issues. "
query_str = (
    "Can you give a high-level overview of BERTScore/MoverScore + formulas if"
    " available?"
)

response0 = query_engine0.query(query_str)
response1 = query_engine1.query(query_str)
response2 = query_engine2.query(query_str)

# query_str = “ROUGE 指标有哪些可重现性问题？提供一些与基准测试相关的详细信息，并描述其他 ROUGE 问题。” query_str = ( “你能否提供 BERTScore/MoverScore 的高级概述 + 如果有公式，也一并提供？” )

输入 [ ]

display_response(
    response0, source_length=1000, show_source=True, show_source_metadata=True
)

display_response( response0, source_length=1000, show_source=True, show_source_metadata=True )

输入 [ ]

display_response(
    response1, source_length=1000, show_source=True, show_source_metadata=True
)

display_response( response1, source_length=1000, show_source=True, show_source_metadata=True )

输入 [ ]

display_response(
    response2, source_length=1000, show_source=True, show_source_metadata=True
)

display_response( response2, source_length=1000, show_source=True, show_source_metadata=True )

输入 [ ]

response1.source_nodes[0].node.metadata

response1.source_nodes[0].node.metadata

输出[ ]

{'questions_this_excerpt_can_answer': '1. What is the advantage of using BERTScore over simpler metrics like BLEU and ROUGE?\n2. How does MoverScore differ from BERTScore in terms of token matching?\n3. What tasks have shown better correlation with BERTScore, such as image captioning and machine translation?'}